Chart Mining-based Lexical Acquisition with Precision Grammars

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 10–18, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Chart Mining-based Lexical Acquisition with Precision Grammars Yi Zhang,♠ Timothy Baldwin,♥♦ Valia Kordoni,♠ David Martinez♦ and Jeremy Nicholson♥♦ ♠ DFKI GmbH and Dept of Computational Linguistics, Saarland University, Germany ♥ Dept of Computer Science and Software Engineering, University of Melbourne, Australia ♦ NICTA Victoria Research Laboratory yzhang@coli.uni-sb.de, tb@ldwin.net, kordoni@dfki.de, {davidm,jeremymn}@csse.unimelb.edu.au Abstract In this paper, we present an innovative chart mining technique for improving parse cover- age based on partial parse outputs from preci- sion grammars. The general approach of min- ing features from partial analyses is applica- ble to a range of lexical acquisition tasks, and is particularly suited to domain-specific lexi- cal tuning and lexical acquisition using low- coverage grammars. As an illustration of the functionality of our proposed technique, we develop a lexical acquisition model for En- glish verb particle constructions which oper- ates over unlexicalised features mined from a partial parsing chart. The proposed tech- nique is shown to outperform a state-of-the-art parser over the target task, despite being based on relatively simplistic features. 1 Introduction Parsing with precision grammars is increasingly achieving broad coverage over open-domain texts for a range of constraint-based frameworks (e.g., TAG, LFG, HPSG and CCG), and is being used in real-world applications including information ex- traction, question answering, grammar checking and machine translation (Uszkoreit, 2002; Oepen et al., 2004; Frank et al., 2006; Zhang and Kordoni, 2008; MacKinlay et al., 2009). In this context, a “preci- sion grammar” is a grammar which has been engi- neered to model grammaticality, and contrasts with a treebank-induced grammar, for example. Inevitably, however, such applications demand complete parsing outputs, based on the assumption that the text under investigation will be completely analysable by the grammar. As precision grammars generally make strong assumptions about complete lexical coverage and grammaticality of the input, their utility is limited over noisy or domain-specific data. This lack of complete coverage can make parsing with precision grammars less attractive than parsing with shallower methods. One technique that has been successfully applied to improve parser and grammar coverage over a given corpus is error mining (van Noord, 2004; de Kok et al., 2009), whereby n-grams with low “parsability” are gathered from the large-scale out- put of a parser as an indication of parser or (pre- cision) grammar errors. However, error mining is very much oriented towards grammar engineering: its results are a mixture of different (mistreated) lin- guistic phenomena together with engineering errors for the grammar engineer to work through and act upon. Additionally, it generally does not provide any insight into the cause of the parser failure, and it is difficult to identify specific language phenomena from the output. In this paper, we instead propose a chart min- ing technique that works on intermediate parsing re- sults from a parsing chart. In essence, the method analyses the validity of different analyses for words or constructions …

Content truncated to 500 words